Try Hanzo
Back to Blog
Hanzo Gen
October 18, 2026
3 min read

A New Pareto Frontier

Better, cheaper, faster: pick all three. Hanzo Gen weaves Enso and Kai into the coding agent you already use, so every step costs what it needs to and no more.

The harnessGenRuns Claude Code, Codex, Cursor or Dev with Hanzo’s tools attached and its model calls on Hanzo.
The routerEnsoSends each call to the model that answers it right for the least, and a hard one to the best.
The deciderKaiAnswers the agent’s choices with a probability for each option. Writes no text.

The trade nobody should make

Every model sits on a curve: a better answer costs more and takes longer. A coding agent pays that curve on every step, and most of its steps never needed the top of it. Reading a file, choosing a tool, deciding a test passed: the frontier model does all of it at frontier prices.Moving the curve takes two things: put each request on the model it actually needs, and stop generating text where a decision will do.

Enso picks the model

Enso is a router trained on outcomes. Each request goes to the model that answers it right for the least, and a hard one goes to the best there is. Its tiers are prices, Flash, Pro and Ultra, and the routing under each is learned. Route statically, by your own policy, or train it on your own traffic as we train ours.
Reading Enso's research runs from api.hanzo.ai
GPQA Diamond accuracy against what 1,000 questions cost, drawn from the research runs as this page loads.

Kai makes the small calls

Most of an agent's steps are choices: run this command or ask first, did the test pass, is the task done. Kai answers each as a decision, a probability for every option, in sub-millisecond latency on GPU (0.27 ms p50). It writes no text, so no output is billed, and its input costs $0.021 per million tokens.

Gen weaves them into your agent

Gen is the harness. It launches the agent you name, Claude Code, Codex, Cursor or Dev, attaches the Hanzo MCP toolset, links the session to your account and sends the agent's model calls through api.hanzo.ai, where Enso routes them. Dev, the agent Gen opens when you name none, runs with Kai deciding each step.Same agent, same workflow. The curve moves under it.
curl -fsSL https://hanzo.sh | sh
hanzo dev --claude
Start with the agent you have